Analyzing large collections of police audio recordings for key terms usually requires human transcribers, a process that is slow and costly. A new arXiv paper highlights this bottleneck and points to off-the-shelf automatic speech recognition (ASR) as an inexpensive, scalable alternative.
The catch is that off-the-shelf ASR output is not fully trustworthy. The paper's title focuses on unreliable ASR, and the abstract points to systematic limitations that can make straightforward keyword counts misleading. The authors therefore tackle the problem of estimating keywords reliably despite this imperfect input.
The abstract does not yet detail the specific methods or results. What is clear is the motivation: making it feasible to mine large police audio datasets without the expense of complete human transcription.